claude-spells

Claude Spells

Claude Code plugins for Java/JVM performance engineering — JFR analysis, flamegraph profiling, and benchmark reporting.

Personal Claude Code plugins for everyday tooling.

Installation

Plugins must be explicitly registered. Run the install script once to add this repo as a marketplace, then install individual plugins:

# 1. Register the marketplace
./install.sh

# 2. Install a plugin (inside Claude Code)
/plugin install jfr-analyzer@claude-spells

Requires jq (brew install jq).

Structure

Each plugin lives in its own subdirectory and follows the Claude Code plugin format:

<plugin-name>/
├── .claude-plugin/
│   └── plugin.json       # Plugin metadata
├── commands/             # Slash commands
│   └── <command>.md
└── skills/               # Skills triggered by context
    └── <skill-name>/
        ├── SKILL.md
        ├── scripts/
        └── references/

Plugins

jfr-analyzer

Analyze Java Flight Recorder (JFR) text exports for performance bottlenecks. Also supports recording JFR profiles directly from live JVM processes.

Analyze an existing file:

• Command: /jfr [file-path] [src=/path/to/src] [lang=zh|en]
• Analyzes: GC pauses, CPU hotspots, memory allocation, thread activity, lock contention, I/O latency
• Input: .jfr binary files (auto-converted) or plain-text exports
• Output: Two Markdown documents — JFR technical report + project optimization plan

Record from a live JVM process:

• Command: /jfr-record [pid=<pid>] [duration=60s] [output=/tmp/app] [analyze=true|false] [src=/path/to/src] [lang=zh|en]
• Requires: JDK with jcmd in PATH (JDK 11+)
• Flow: lists Java processes → starts jcmd JFR.start → waits → converts .jfr → auto-analyzes
• Default: duration=60s, analyze=true, lang=zh

End-to-end example:

/jfr-record pid=12345 duration=120s src=~/projects/myapp lang=zh

Records 120s of profiling data from PID 12345, then produces a full Chinese-language analysis report.

See jfr-analyzer/ for details.

esrally-analyze

Analyze esrally benchmark reports and generate structured technical summaries.

• Skill: invoke by providing .md, .tar.gz, or .gz rally result files, or ask to "分析压测结果" / "compare benchmark runs"
• Supports: single-run analysis, A/B comparison (e.g. different translog/indexing configs), multi-run baseline statistics
• Output: structured Markdown report with throughput, latency, and per-operation breakdowns
• Languages: Chinese (lang=zh) and English (lang=en)

See esrally-analyze/ for details.

flamegraph-analyzer

Analyze async-profiler HTML flamegraph files for any JVM application.

• Skill: invoke by providing a .html flamegraph file, or ask to "分析火焰图" / "analyze flamegraph"
• Framework detection: auto-detects Spring, Kafka, Flink, Elasticsearch, Lucene, Netty, gRPC, and more
• Analyzes: CPU hotspots, layer breakdowns, critical call chains, GC/safepoint/lock contention patterns
• Focused analysis: use focus=translog,gc or natural language like "重点关注 translog" to spotlight specific topics
• Output: Markdown report with per-layer CPU distribution, top hotspot methods, prioritized recommendations, and optional focused observations section
• Languages: Chinese (lang=zh) and English (lang=en)

See flamegraph-analyzer/ for details.